Ship everything your agents learn.

Evidence first, then policy. A Candidate becomes Verified when deterministic checks back it: the agent's own test run, exit codes, CI. If verifiers disagree, it is not promoted. Trusted asks for more: by default three independent passing runs (five for destructive actions), still fresh and compatible, plus a human approval for high-risk work. A signature proves who made an object, never that it is safe to reuse.

Stale and revoked versions are withheld from relevant local use. A reviewed context profile records selected file and runtime fingerprints; legacy or unobserved context can remain UNKNOWN. Drift scanning and fresh revalidation support review. High-risk actions require current compatibility and policy checks. Revocation writes a tombstone; peers receive it when they sync, so an offline peer can lag.

Captured raw traces stay in your encrypted local vault; sync has no raw-trace upload route. By default the cloud sees control metadata only: IDs, plan and request counters. Searchable Sync adds sanitized step templates and fingerprints; file contents, diffs, command output, prompts and absolute paths stay on your machine. Opaque Relay sends full steps, end-to-end encrypted. Cloud copies show their own cloud_state and reach Trusted only with independent evidence from your devices. Explicit cloud recall sends your query and any selected sanitized failure context; optional BYOK embeddings also send selected text to your model provider. Sync is off until you turn it on under the sync terms:

agentevo login
agentevo sync mode --set searchable --terms-version 2026-09-21 --yes
agentevo sync push
agentevo recall "fix flaky test" --capabilities Bash,Edit,Read --cloud

Claude Code ships today: --latest finds your newest session on its own. Codex is in beta and reads sessions exported to Agentevo's Codex JSONL format, not Codex's own session files yet. Cursor, OpenCode, and CI join them through capture --adapter cursor|opencode|ci, all over Agentevo-defined JSONL contracts — the adapters never read proprietary session stores. QA / security agents are next. Adapters must pass conformance fixtures, and a connector update can never silently raise classification or permissions. The local guard targets p95 under 50ms and works offline. Compilation and verification are separate operations; configured host hooks and selected test replay can add session latency.

No. Model inference is bring-your-own-key by default: your agent keeps using its own provider and keys. The honest boundary: local-first governs data sent to Agentevo. Your agent's own calls to its model provider are a separate disclosure.

On Free, counted cloud operations pause and everything local keeps working: capture, guard and the registry. Cloud recall pauses too — re-run with explicit --offline to search this machine's index instead. Export, deletion and revocation remain available at quota. Free never bills overages. Paid-plan limits and billing options will be published when available.

Yes. Capture, compile, local verification, promotion, guard, and the registry work locally. Initialized projects recall locally by default; --offline also selects local recall outside them. Explicit --cloud needs Searchable Sync; Opaque Relay transfers encrypted objects for local indexing and recall after pull. Customer-controlled deployment can be scoped in an enterprise pilot. Public-cloud sync is opt-in and not required for the core learning loop.

Enterprise pilots can scope customer-controlled deployment and additional controls. Availability is agreed per pilot; the local guard still applies. Destructive actions require the configured evidence and approval gates.